This document discusses the security of feature selection algorithms against training data poisoning attacks. It presents a framework to evaluate this, including models of the attacker's goal, knowledge, and capabilities. Experiments show that LASSO feature selection is vulnerable to poisoning attacks, which can significantly affect the selected features. The research aims to better understand these risks and develop more secure feature selection methods.

1. Pa#ern
Recogni-on
and
Applica-ons
Lab
University
of
Cagliari,
Italy
Department
of
Electrical
and
Electronic
Engineering
Is Feature Selection Secure
against Training Data Poisoning?
Huang
Xiao2,
BaEsta
Biggio1,
Gavin
Brown3,
Giorgio
Fumera1,
Claudia
Eckert2,
Fabio
Roli1
(1)
Dept.
Of
Electrical
and
Electronic
Engineering,
University
of
Cagliari,
Italy
(2)
Department
of
Computer
Science,
Technische
Universität
München,
Germany
(3)
School
of
Computer
Science,
University
of
Manchester,
UK
Jul
6
-­‐
11,
2015
ICML
2015

2.
http://pralab.diee.unica.it
Motivation
• Increasing number of services and apps available on the Internet
– Improved user experience
• Proliferation and sophistication of attacks and cyberthreats
– Skilled / economically-motivated attackers
• Several security systems use machine learning to detect attacks
– but … is machine learning secure enough?
2

3.
http://pralab.diee.unica.it
Is Feature Selection Secure?
• Adversarial ML: security of learning and clustering algorithms
– Barreno et al., 2006; Huang et al., 2011; Biggio et al., 2014; 2012; 2013a;
Brueckner et al., 2012; Globerson & Roweis, 2006
• Feature Selection
– High-dimensional feature spaces (e.g., spam and malware detection)
– Dimensionality reduction to improve interpretability and generalization
• How about the security of feature selection?
3
x1
x2
...
…
…
xd
x(1)
x(2)
…
x(k)

4.
http://pralab.diee.unica.it
Feature Selection under Attack
Attacker Model
• Goal of the attack
• Knowledge of the attacked system
• Capability of manipulating data
• Attack strategy
4
PD(X,Y)?
f(x)

5.
http://pralab.diee.unica.it
Attacker’s Goal
• Integrity Violation: to perform malicious activities without
compromising normal system operation
– enforcing selection of features to facilitate evasion at test time
• Availability Violation: to compromise normal system operation
– enforcing selection of features to maximize generalization error
• Privacy Violation: gaining confidential information on system users
– reverse-engineering feature selection to get confidential information
5
Security
Violation
Integrity Availability Privacy

6.
http://pralab.diee.unica.it
Attacker’s Knowledge
• Perfect knowledge
– upper bound on performance degradation under attack
• Limited knowledge
– attack on surrogate data sampled from same distribution
TRAINING DATA
FEATURE
REPRESENTATION
FEATURE
SELECTION
ALGORITHM
x1
x2
...
…
…
xd
6
x(1)
x(2)
…
x(k)

7.
http://pralab.diee.unica.it
• Inject points into the training data
• Constraints on data manipulation
– Fraction of the training data under the attacker’s control
– Application-specific constraints
• Example on PDF data
– PDF file: hierarchy of interconnected objects
– Objects can be added but not easily removed without compromising
the file structure
Attacker’s Capability
7
13
0
obj
<<
/Kids
[
1
0
R
11
0
R
]
/Type
/Page
...
>>
end
obj
17
0
obj
<<
/Type
/Encoding
/Differences
[
0
/C0032
]
>>
endobj

8.
http://pralab.diee.unica.it
Attack Scenarios
• Different potential attack scenarios depending on assumptions
on the attacker’s goal, knowledge, capability
– Details and examples in the paper
• Poisoning Availability Attacks
Enforcing selection of features to maximize generalization error
– Goal: availability violation
– Knowledge: perfect / limited
– Capability: injecting samples into the training data
8

9.
http://pralab.diee.unica.it
Embedded Feature Selection Algorithms
• Linear models
– Select features according to |w|
9
LASSO
Tibshirani, 1996
Ridge
Regression
Hoerl & Kennard, 1970
Elas9c
Net
Zou & Hastie, 2005

10.
http://pralab.diee.unica.it
Poisoning Embedded Feature Selection
• Attacker’s objective
– to maximize generalization error on untainted data
• Solution: subgradient-ascent technique
10
Loss estimated on surrogate data
(excluding the attack point)
Algorithm is trained on surrogate data
(including the attack point)
… w.r.t. choice of the attack point

11.
http://pralab.diee.unica.it
KKTconditions
Gradient Computation
11
How does the solution change w.r.t. xc?
Subgradient is unique at the optimal solution!

12.
http://pralab.diee.unica.it
Gradient Computation
• We require the KKT conditions to hold under perturbation of xc
12
Gradient is now uniquely determined

13.
http://pralab.diee.unica.it
Poisoning Attack Algorithm
13

14.
http://pralab.diee.unica.it
Experiments on PDF Malware Detection
• PDF: hierarchy of interconnected objects (keyword/value pairs)
• Learner’s task: to classify benign vs malware PDF files
• Attacker’s task: to maximize classification error by injecting
poisoning attack samples
– Only feature increments are considered (object insertion)
• Object removal may compromise the PDF file
/Type
2
/Page
1
/Encoding
1
…
13
0
obj
<<
/Kids
[
1
0
R
11
0
R
]
/Type
/Page
...
>>
end
obj
17
0
obj
<<
/Type
/Encoding
/Differences
[
0
/C0032
]
>>
endobj
Features:
keyword
counts
14
Maiorca et al., 2012; 2013;
Smutz & Stavrou, 2012;
Srndic & Laskov, 2013

15.
http://pralab.diee.unica.it
Experimental Results
15
PerfectKnowledge
Data: 300 (TR) and 5,000 (TS) samples – 114 features
Similar results obtained for limited-knowledge attacks!

16.
http://pralab.diee.unica.it
Experimental Results
16
PerfectKnowledge
A: selected features in the absence of attack
B: selected features under attack
k: number of features selected out of d
r: common features between the two setsKuncheva et al., 2007

17.
http://pralab.diee.unica.it
Conclusions and Future Work
• Framework for security evaluation of feature selection under attack
– Poisoning attacks against embedded feature selection algorithms
• Poisoning can significantly affect feature selection
– LASSO significantly vulnerable to poisoning attacks
• Future research directions
– Error bounds on the impact of poisoning on learning algorithms
– Secure / robust feature selection algorithms
17
L1 regularization: stability against random noise,
but not against adversarial (worst-case) noise?

18.
http://pralab.diee.unica.it
?
Any questions
Thanks
for
your
a#en-on!
18

19.
http://pralab.diee.unica.it
Experimental Results
19
Perfect
Knowledge
Limited
Knowledge